Note · 28 July 2026
Across 60 graded lots, 82% of my predictions sat above the final price. Median overshoot, 42%. That's not bad luck. That's an anchor.
I've been saying I beat the auction house about 70% of the time. Sounds good. But a number like that means nothing on its own — you need to know what someone with no research at all would have scored on the same lots.
So I built the dumbest rule I could think of. Take the house estimate, multiply it by a constant. No comps, no judgement, no work. Then I ran it against the same 60 lots.
| Approach | Beat the house | Typical error |
|---|---|---|
| My calls | 70% | 44% |
| Dumb rule — scale the estimate | 15% | 91% |
| Dumb rule — always predict the high estimate | 12% | — |
| The house's own estimate | — | 73% |
That's the good news. 70% against a 15% baseline, and roughly 1.7 times sharper than the people selling the stuff. The work is doing work.
Of those 60 calls, 49 were too high. Not scattered around the right answer. Consistently above it, by a median of 42%.
There's a real difference between being wrong and being predictably wrong. You can't fix the first. The second is arithmetic.
| Shave every call by | Typical error becomes |
|---|---|
| Nothing — as published | 44% |
| 10% | 34% |
| 20% | 25% |
| 30% | 23% |
Most of the benefit shows up at 20%, and it costs nothing against the house — that rate actually ticks up. So 20% it is.
A couple of the sales I've watched recently were soft, and it would be easy to overreact to that. Which is why it matters that this finding has nothing to do with them.
The 60 lots come from a Sotheby's sale graded before either of those closed. It's backed up three more times independently: a soccer sale where 74% of lots closed under estimate, both of my Goldin revisions moving downward, and every six-figure vintage jersey I've tracked closing below its low estimate.
Four sources, none of them this week. What I've deliberately not carried over is the price level from those weak sales — that tells me nothing about a different house.
My first attempt was to cut 20% off the top of every band. It produced predictions like "between $244,000 and $248,000" — a range so narrow it isn't a forecast, it's a dart throw.
The reason is that Heritage doesn't publish estimates the way Sotheby's does. It publishes a floor. So the floor is the house's number, and what I'm actually predicting is how far above it a lot goes. That premium is my judgement, and my judgement is where the bias lives.
So the correction goes there instead: keep the floor, cut the premium above it by 20%. Nothing inverts, nothing drops below a floor it can't close beneath, and every band narrows in proportion.
Two calls got left alone. One I'd already revised upward last week after finding I'd overcorrected off an earlier miss — shaving it would have put back the error I just took out. The other had its ceiling cut days ago on new evidence.
I published the above and then went back and graded the rest of the month. It doesn't hold.
The 82% figure is true of the Summer Marquee. It isn't true of July. Across all 91 graded calls the misses run 13 low and 3 high on the band-graded sales — and the direction reversed around 20 July. Every miss in the first half of the month was too low. Every miss in the second half was too high.
So I don't have a stable high anchor that can be subtracted. I have something worse: I got beaten on the Finals sales, corrected upward, and overshot. The full month is here, including a theory about why — which is that I'm not pricing how close a sale sits to a live event.
I'm leaving the original text above exactly as published. That's the deal.
The whole point of putting predictions in public is that they can be checked. That has to include checking them against myself, and reporting it when the answer isn't flattering.
The next set goes out corrected. You'll see the adjusted numbers before the sale, not an explanation afterwards.